The Research And Implementation Of Face Detection And Tracking Based On Adaboost And TLD

The research on facial image processinghas attracted increasing attention as one of theimportant topics in the field of computer vision. The processing of face imageis related tocomputer graphics, computer vision, pattern recognition, machine learning, cognitive science,artificial intelligence, computing intelligence and a variety of other technologies. Currentresearch on face image processing has focused on the face detection, face tracking, facerecognition, facial expression recognition, gesture analysis, face synthesis and several otherareas. This series of studies are based on the research results of face detection and tracking. Atpresent, research and tracking of human face has been widely used in the field of videosurveillance and human-computer interaction. This paper focuses on face detection and facetracking.This paper firstly studies the problem of static image Human Face Detection, includingthe most popular cascaded detection algorithm based on Adaboost algorithm. It describes theconstruction of Haar features, integral image stratery, selection of weak classifiers, as well ashow to train a cascaded classifier. The study shows that the Adaboost algorithm trainingprocess is very time consuming. Although the final constructed cascade classifier can detectface images rapidly, but the detection rate also decreases correspondingly. Secondly, bycombining Haar features and MB-LBP features, this paper proposes a new image descriptionoperator, which can accurately describe grayscale variationsof local image information. Theexperiments indicate that the classification capability of this new operator is more powerfulthan Haar features and MB-LBP features. By adopting the new operator as weak classifiersand utilizing the RealAdaboost algorithm, the training of cascaded classifier can beaccomplished with less number of features while achieving the same detection effect as Haarfeatures. At the same time, the training time is also efficiently reduced since the BitBP featurehas a much smaller size. Combined with the characteristics of the new operator, this paperpresents a new multiple cascaded classifier method, which ensures the detection results of theclassifier while effectively simplifying the process of training. Finally, in the field of facetracking, we study and improve the current popular Tracking-Learning-Detection (TLD)algorithm which divides the track problem into three separate but related modules that caneffectively overcome the occlusion, the target disappeared and repositioning and other issues. This paper combined the Mean Shift tracking algorithm and TLD framework and built a newLTD tracking system that can achieve the rotation as well as fast-moving target tracking.Finally, this paper combines the Mean Shift based TLD framework with face detector to builda tracking framework to address the face detection and tracking problem. Experimental resultsdemonstrate that the proposed algorithm shows4.5%higher recall rate and9.8%higherprecision rate than TLD when tracking the out-of-plane rotation face.